Rich Barton-Cooper and Adam Gleave argue that researchers lack safe channels to report dangerous jailbreaks, methods for circumventing AI safeguards, to frontier model developers. They note that extremist organizations have reportedly trained members on jailbreaking techniques and that universal jailbreaks have been found across advanced models. The piece proposes immediate reforms including appropriately scoped non-disclosure agreements, publicly disclosed severity rubrics, and year-round disclosure pathways. For a longer-term solution, the authors advocate establishing an independent third-party clearinghouse modeled after cybersecurity’s CERT/CC, which would accept vulnerability reports, grade them consistently, coordinate fixes across affected labs, and publish aggregate security metrics.
